Representations & Consciousness: Chapters 4 & 5

Question 1

What are examples of conscious modes?

Answer 1

perception, imagery, dreaming

Question 2

What are the problem with direct definitions of consciousness?

Answer 2

Direct definitions of C usually resort to circularity (“seeing”, “knowing”, “realising”)

Question 3

What is a better approach than defining consciousness head on?

Answer 3

Instead of defining C head-on, we better start with asking:
What can be peeled away from our lives before C is lost?

Question 4

Several modalities have been investigated as necessary components. Which of these does Pennartz posit empirical support has been found for consciousness in the absence of?

Answer 4

Empirical support for conscious experience
in the absence of:
-Motor activity
-Language (incl. verbal beliefs, judgment)
-Emotion
-Memory

Question 5

How can conscious vision be broken down?

Answer 5

Conscious vision can be broken down into
various components, e.g.:
- Color vision (involving V4 & inferotemporal patches)
- Motion vision (MT/V5)
- Form and Face vision (FFA, IT etc.)
- Vision in an entire hemifield (parietal cortex - hemineglect)

Question 6

Why is breaking down vision like this relevant?

Answer 6

Important to look at what can’t be peeled away:
MT/V5 –> if away –> akinetopsia
loss of V4 –> achromatopsia

what can’t be peeled away?
:: pieces of sensory ctx
:: parietal ctx
:: Thalamocortical systems

Consciousness in other modalities also depends on specific cortical systems (e.g. auditory, somatosensory cortex,
olfaction, taste)

Question 7

What are Pennartz’ hallmarks of consciousness?

Answer 7

Hallmarks:
* Qualitative (multimodal) richness
* Situatedness & immersiveness: you’re right in the middle of the situation (immersed into it)
* Integration, unity
* Dynamics and stability: establishing of objects
* Interpretation, inference, intentionality

Question 8

How does Pennartz construct his definition of consciousness?

Answer 8

Modes of (healthy) consciousness: perception imagery, dreaming
Hallmarks:
* Qualitative (multimodal) richness
* Situatedness & immersiveness: you’re right in the middle of the situation (immersed into it)
* Integration, unity
* Dynamics and stability: establishing of objects
* Interpretation, inference, intentionality

Definition of conscious experience: Inferential representation that is situational (spatially encompassing) and multimodally rich

Question 9

Describe the hard problem of cosciousness

Answer 9

Past decades: much progress on “easy” problems of
consciousness – memory, attention, decision-making, sensory
discrimination (etc.)
* But: we refrain from asking deeper questions, e.g. how is
phenomenal content associated with neural activity (”hard”
problem; D. Chalmers; “Explanatory Gap” - Levine)
* What is phenomenal content?
– having qualitatively rich experiences
– What is it like to be…. (e.g. you)?

Question 10

Give two examples of the difference between easy and hard problems

Answer 10

• First example: painting by Van Gogh àcorrelate pictorial
  elements (shape, color, etc.) with neural activity in different
  brain areas; but what is their exact relationship?
• Second example: physical description of light vs. color experience

Question 11

What is a key problem of consciousness?

Answer 11

“whole-pattern perception” (Hallmark: integration, unity)

Question 12

What group of psychologists tackled this problem of whole pattern perception?

Answer 12

This is a classic problem in Gestalt Psychology: e.g. Kurt Koffka, Max Wertheimer and Wolfgang Kohler
=> “Holistic” vs. analytic approach to perception

Question 13

How do we distinguish whole objects against a background according to Gestalt psychologists?

Answer 13

Gestalt psychologists: whole-figure recognition explained from common features present in parts of the figure

Gestalt “laws” of perception:
* Law of common fate (common motion)
* Law of good continuation
* Law of similarity, proximity, closure

Question 14

What neural network is compared to these Gestalt psychologists and why?

Answer 14

Gestalt Psychology and whole-pattern recognition by recurrent nets: Gestalt laws suggest how bottom-up grouping of features
into an object may be achieved. Attractor properties of recurrent nets may help explain bistable (flip/flop, 2 basins) nature

Question 15

Do Gestalt laws therefore explain holistic perception?

Answer 15

Gestalt laws explain feature grouping, not the holistic (all-or-none) aspect of perception

Question 16

What can neural network models explain about cognition? (2)

Answer 16

Stability of percepts (and/or imagery) is a hallmark of
(conscious) representation – also achieved in recurrent nets

Emergence: networks show how low-level phenomena can
give rise to more complex, high-level phenomena (Imagine you are a neuron connected to a large array of neurons: no clue what you and others are representing (but the representation is there, “supra-neural”))

Question 17

What did the success of neural network theory lead to?

Answer 17

Success of neural network theory led to neurocomputational
account of consciousness

Question 18

Give two examples of these neurocomputational accounts of consciousness

Answer 18

*e.g. Paul and Patricia Churchland’s eliminative materialism
attempts to explain away all mental phenomena by brain’s
physical processes (eliminate “folk psychology”)

*e.g. Explain memory, attention, multistability (etc.) from
recurrent properties in corticothalamic systems

Question 19

How is this recurrent processing of consciousness described in P. M. Churchland (1995)?

Answer 19

The thalamocortical loop is the posited recurrent properties responsible or involved in consciousness; the intralaminar nuclei especially have far reaching ascending and descending pathways around the cortex.

Question 20

What caveats are pointed out about this recurrent processing theory?

Answer 20

*Intralaminar nuclei project less specifically than depicted

*Intralaminar nuclei are important for arousal, not modality-specific recurrent processing

*Recurrent pathways found elsewhere in the brain, also in “nonconscious” structures such as cerebellum.

Question 21

How do Patricia Churchland and Terry Sejnowski approach the question of how neurons represent anything?

Answer 21

“…. in view of the opportunity to correlate neuronal
responses with controlled stimuli, the sensory systems are
a more fruitful starting point for addressing this question
than, say, the more centrally located structures such as the
cerebellum or the hippocampus or prefrontal cortex. […]
Constrained by transducer output, the brain builds a model
of the world it inhabits.

That is, brains are world-modelers, and the verifying
threads —the minute feed-in points for the brain’s
voracious intake of world-information — are the neuronal
transducers in the various sensory systems.”

Question 22

How do Patricia Churchland and Terry Sejnowski approach the question of how neurons represent anything?

Answer 22

“…. in view of the opportunity to correlate neuronal
responses with controlled stimuli, the sensory systems are
a more fruitful starting point for addressing this question
than, say, the more centrally located structures such as the
cerebellum or the hippocampus or prefrontal cortex. […]
Constrained by transducer output, the brain builds a model
of the world it inhabits.

That is, brains are world-modelers, and the verifying
threads —the minute feed-in points for the brain’s
voracious intake of world-information — are the neuronal
transducers in the various sensory systems.”

Question 23

What omission does Pennartz point out in this account?

Answer 23

But: How could sensory receptors act to verify that a model of the world (~ representation) is correct? (What is the independent evidence?)

In other words: inputs & outputs of neural nets are not specified / identified
(except by external observer)

=>the network itself has “no clue” about what it is processing (it processes numbers)

Question 24

Why is the input to our perceptual systems not specified? There is no shortage of feature detectors within each modality (submodalities) so that’s not the problem. Name two classic hypotheses on this issue

Answer 24

First classic hypothesis: pattern coding
Second hypothesis: labeled-lines coding

Question 25

Describe pattern coding

Answer 25

Suppose the brain needs to be informed about two taste inputs, ‘bitter’ and ‘sweet’: Different chemical applied to same taste bud could result in different sequences of activations encoding the two different inputs

Different types of information sent across 1 common ‘line’
(=receptor/nerve fiber)

Question 26

What was the fate of this pattern coding hypothesis?

Answer 26

This hypothesis was refuted in the 19th century

Question 27

Describe labelled lines coding

Answer 27

It is a specific receptor type/nerve fiber that conveys the modality and the kind of signal propagated is the same for different modalities

Question 28

How was this labelled-lines coding received?

Answer 28

This hypothesis received widespread confirmation

Question 29

What was included in Müller’s doctrine of Specific Nerve Energies (1838)?

Answer 29

Ten laws, including:
» Law 3: “The same external cause [..] gives rise to different sensations in each sense, according to the special endowments of its nerve“ (Not important how it is activated so long that it is)

> > Law 5: “We are directly aware, not of objects, but of activity of our nerves themselves“ (not aware of APs but of results/consequence of that AP )

Law 3: e.g. Smack on eye àtactile sensation but also visual percept

Question 30

How could these laws be interpreted/ evaluated now?

Answer 30

What are the “special endowments of nerves”?
=>We now know: different nerves conduct action potentials in a
very similar way

How can we be aware of activity of our nerves themselves?
=> “activity of our nerves” would have to be: spike trains
Alternatively: how does the brain convert spike trains into
percepts?

Question 31

Critically evaluate this labelled lines hypothesis

Answer 31

From the brain’s viewpoint, the labeled-lines hypothesis is cheating: anatomical origin of afferent input is unknown to it. Brain area X receives spike trains / EPSPs, but these do not disclose the identity of the source

All types of nerve fibers operate in the same way: propagation of action potentials to brain
=> the problem of modality ́labelling ́ (or: identification) has been relegated to the brain

=> This problem becomes the brain’s problem of Modality Identification: how does the brain know what kind of information it is processing?

Question 32

Describe a counter argument this these criticisms of labelled lines hypothesis

Answer 32

Argument: Visual input is differently structured than auditory input because of the way the eyes and retina are constructed (spatial field; disparity) and photon patterns impinge on it

Question 33

What is a criticism of this structural explanation of the labelled lines problem?

Answer 33

But: the ‘visual’ face pattern in an array of neurons could also represent an activity pattern on a tonotopic or somatosensory map; it is not inherently visual

And: dynamic properties may refer to visual movement, changes in pitch, stroke across the body

So: Answer is No, local input statistics do not offer a full explanation

Question 34

What is the argument for feature detectors playing a role in this?

Answer 34

Argument: each sensory modality is specific (identifiable) by virtue of detectors that are uniquely present only in that modality (e.g. hearing – consonance, harmony)

Question 35

Give Pennartz’ view on this feature detector argument

Answer 35

-Imagine unique feature detectors in the brain for vision and taste, but unconnected. Formally (topologically) it does not matter if these are separated physically, even across two brains; so how would a brain tell a sensory input is unique?

An isolated feature detector responds to a specific stimulus, but so does a smoke detector (non-conscious). Why would an unconnected feature detector code any particular modality / sensation? So: Answer is No: unique feature detectors alone do not help us out

Question 36

What does Searle, based on his Chinese room, say that brains must have that computers do not?

Answer 36

The brain must have causal powers (lacking in computers) that endow it with consciousness

Question 37

The Chinese room poses a problem for computers having consciousness, but what does it mean for the brain?

Answer 37

the brain does have a ‘mind’ and also lives in a kind of Chinese room: the Cuneiform room. Meaning of inputs and outputs is also unknown to the brain – they are unidentified spike patterns

The Chinese room is a legitimate problem, but it equally applies to the brain (not just to computers). Using a diversity of (unidentified) sensory inputs, the brain must construct a world model that identifies (or infers) modalities by itself

Question 38

What could Searles ‘causal powers’ be or not be?

Answer 38

Searle’s “causal powers” of the brain: unlikely to be
new physicochemical properties (~ vitalism)

We must rather search for unique organisational /computational principles

Question 39

How are neural network models limited in cognitive neuroscience?

Answer 39

Neural network models: suffice to explain some mental properties, but do not solve (e.g.) Modality Identification problem (=> problem of meaning)

Question 40

If we compare neural nets to non-living things in nature, what are the possible outcomes?

Answer 40

–neural networks could be unique: would confirm validity of neurocomputational approach

–if inanimates show similar functions: are neural nets too limited, or underconstrained?

–inanimates showing similar functions: this could also mean that ‘mind’ is present in many inanimate systems => panpsychism

Question 41

Give an example of what an inanimate neural network could look like. What does Pennartz conclude from this?

Answer 41

A “Rocky” neural network:
* Sunlight provides “sensory input”
* Neural connections: infrared radiation from rock to rock
* Activation state: rock’s temperature
* Output of neural net: classify sensory input as “cool…..luke-warm….hot”
* Task may be too simple: but also direction of sunlight can be decoded from network outputs
* Conclusion: basic properties of neural networks fulfilled by collection of rocks

Question 42

Maybe there is recurrent feedback lacking? Does it lack learning and memory capacity? How does Pennartz respond to these?

Answer 42

Possible objection: recurrent feedback lacking?
(cf. P. Churchland)
=> can be fixed by particular arrangement of rocks

*Does the rock formation lack learning and
memory capacity?
=> storage of heat (kinetic energy) counts as
memory?
=> also: pattern of rock-rock connectivity can be modified when light input changes (e.g. clouds can induce “synaptic plasticity”)

*Does a sandy beach store a memory of your foot?
=> network of cohesive (connected) sand particles
*If beach would have pattern-completing abilities, would you consider the beach conscious?

Question 43

What standpoint is the chinese room thought experiment against?

Answer 43

Against hard functionalist standpoint of ‘computers are a good metaphor for the mind/brain’

Question 44

Should we thus accept panpsychism? What is Pennartz’ objections to this? (3)

Answer 44

:: Overt behaviour is lacking => not decisive
:: Rocky networks lack representations (dreaming, perceiving, imagining, thinking-in-language)
:: Beach may have “passive” representation (=mold), but this contrasts with memory
–as active re-creation and reconstruction
–as something intended to be used as memory

Question 45

What therefore is Pennartz’ take on panpsychism?

Answer 45

Hard to reject panpsychism completely, but: alternative is more likely, viz. to classify (classic) network models as insufficient / under-constrained when one aims to explain consciousness

Question 46

In what sense are neural networks underconstrained?

Answer 46

Models do not solve MI problem (=> meaning)

There exist brain systems structured as neural nets, yet are not linked to consciousness (e.g. cerebellum)

Brain systems that are associated with consciousness, can become ‘unconscious’ (e.g. anesthesia) but still behave like neural nets then

Question 47

What would panpsychism do to terms like consciousness?

Answer 47

Panpsychism would lead to an overgeneralised use of terms like Consciousness => meaningless term; why not just say complexity

Question 48

How does Pennartz relate consciousness to physics?

Answer 48

Consciousness refers to certain objects showing manifest properties (of being conscious)
=> compare it to magnetism or frozenness (in physics)

Question 49

Describe global workspace theory

Answer 49

Global workspace is a hub or informational ‘marketplace’ from which information is broadcast. Broadcasting is mainly via corticocortical fibers. Pre-conscious vs conscious stages of processing involves crossing a threshold => “ignition” process

Question 50

What are critiques of GNWT? (4)

Answer 50

–model addresses logistics of information exchange and “access” aspects
–does not address phenomenal content
–implementations are classic neural network models, facing the same problems
–there is evidence for “ignition” (-> PFC), but only if perception is coupled to motor response (reporting)

Question 51

What is meant by this phenomenal vs access consciousness?

Answer 51

Phenomenal vs. Access Consciousness: experienced content vs cognitive processes operating on this content (e.g. reporting, working memory, attention)

Question 52

What theory of consciousness was approached by Tononi?

Answer 52

Tononi: substrate for consciousness is large cluster of neuronal groups (“dynamic core”, coding color, motion, shape,..)

conscious state: high degree of differentiation of information coded
conscious state: high degree of integration (defined by Information Theory)

Question 53

But what would the ‘core’ of information theory be? In terms of neural structures?

Answer 53

Membership of Core is flexible and changes rapidly (members: cells in MT, V4, etc.)

Question 54

How would a camera hold up in IIT?

Answer 54

Digitised camera picture contains differentiated information in terms of pixel activation, but there is no integration

Question 55

How is information defined within information theory?

Answer 55

Information is Reduction of Uncertainty e.g in Guess who, as more questions are asked chance that person p is selected increases. A bit (binary digit) is the amount of information needed to choose between two equally likely alternatives:

Start of game; Maximal uncertainty
2 alternatives: 0 or 1
Add 2 alternatives: 0 or 1; 00,01,10,11
Add 2 alternatives: 0 or 1; 000, 001…

In total 8 alternatives (states) are encoded by making 3 binary
decision steps:
8 = 2^3
Or: I = ^2log K = - sum_K( (1/k) log(1/k) )

With I: Information (also: H, Entropy; log of number of alternatives or states; degrees of freedom in a system) (H = I = 3)

K: Number of states or alternatives to be encoded (K=8)

Question 56

Give a generalised form of this equation for discrete variables

Answer 56

I = ^2log K = - sum_i( (pi) log(pi) )
i: indexes the alternatives (1….K)
pi: probability of alternative i
Negative sum of the probability of each outcome * the log probability

Question 57

What happens as pi decreases?

Answer 57

Note: if pi decreases, the negative log increases
more steeply than that pi decreases!
=> Unlikely events convey more information than likely ones

y = - a log x

logarithmic function so like exponential function along y axis

Question 58

What conclusion does Pennartz draw about the relevancy of these equations

Answer 58

NB, Shannon information is about statistics of relationships in a system (not: meaning)

Question 59

Describe the concept of mutual information in IIT; What does it indicate

Answer 59

MI (A;B)= H(A) + H(B) – H(AB)
A and B are subsystems of the overall system, AB

H(A) is the statistical entropy of subset A
( information entropy is the average amount of information conveyed by an event, when considering all possible outcomes.)

• MI is the surplus of possible states A and B can assume alone but summed up, relative to the states of their combination
• If A and B are fully independent and each alone can assume 2 states => MI = 1 + 1 - 2 = 0
• If A and B are fully dependent and each alone can assume 2 states => MI = 1 + 1 - 1 = 1

Thus, Mutual information indicates Statistical dependence

Question 60

How do you measure integration in IIT?

Answer 60

Int(x) = sum(H(xi) – H(X))
xi : member of a larger neuronal group (X)àsummation across all members of X

Int(X): integration within X

Question 61

How do you measure clustering in IIT? What does this mean?

Answer 61

Cl(X) = Int(X) / MI (X;Y)
Cluster Index (CI): some subsystems are more integrated / statistically dependent than others

= ratio between statistical dependence within group X relative to the statistical dependence between X and the rest of the system Y

Question 62

What is Tononi’s claim about the cluster index?

Answer 62

Tononi: a high Cluster Index would correspond to high degree of consciousness (1998; later on: use of Phi-measure)

Question 63

Critically evaluate Tononi’s Integrated Information Theory of consciousness

Answer 63

Captures a number of relevant consciousness phenomena
e.g. dynamism, combinations of “essential nodes” to form a changeable core. This is a serious attempt to produce axiomatic, quantitative theory accounting for “integration” and “differentiation” in perception

But: boils down to quantification of statistical dependence
in and between subsets of neurons.
:: Integration as “scene construction” not explained (but was original goal)
:: Phenomenal awareness not explained (why qualia?)
:: No intentionality and no ”matching” to the world

Question 64

Where does pennartz claim this theory brings us back to?

Answer 64

Panpsychism: e.g weather systems on earth have patches of high / low stat. independence